This weekend, Linus Torvalds (the creator of the Linux kernel) shared his opinion on the instruction set Intel AVX-512 found in some of these processors.
In the comment he left, Torvalds I mention that «expect the AVX-512 to die a painful death«.
For the father of Linux, AVX-512 instruction set is only beneficial for HPC market and it has real drawbacks. According to him, Intel should stop wasting its resources in new instruction sets to focus on the things that really matter.
US chipmaker Intel recently announced that its upcoming Alder Lake processors will not be compatible with the AVX-512 instruction set.
The AVX-512 is a set of 512-bit extensions to SIMD instructions (Advanced Vector Extensions) 256-bit for x86 instruction set architecture that Intel proposed seven years ago. Support for the AVX-512 debuted on the Intel Xeon Phi x200 processor (Knights Landing) in 2016.
Subsequently, AVX-512 broke through to other offers from the chipmaker, primarily Skylake-SP, Skylake-X, Cannon Lake, and Cascade Lake. Currently, the company's Intel Cooper Lake and Ice Lake processors support certain AVX-512 subsets. Intel is rumored to have confirmed that Tiger Lake will operate the AVX-512 before recently announcing that the next Alder Lakes will not operate it.
For Torvalds, this is all useless and he hopes Intel will go back to basics and restart its process.
"I hope that the AVX-512 dies a painful death, and that Intel begins to solve the real problems instead of trying to create magic instructions and then create benchmarks that they can trust," he stated?
As well wants the chipmaker to focus more on regular code, which is not HPC (high performance computing) or any other unnecessary special case. From what Torvalds says, this is all reminiscent of the days when Intel dominated the chip market, but its FP performance was zero.
“I've said it before and I'll say it again: at the heyday of x86, when Intel laughed at the bank and killed their entire competition, absolutely everyone did better than Intel when it came to FP Charges.
Intel's FP performance was zero (relatively speaking), and that didn't matter because, in fact, no one cared outside of the benchmarks. The same goes for the AVX-512 today and in the future. Yes, we can find things that matter. No, these things don't sell machines in general, ”he said.
Linus Torvalds thinks the AVX-512 has real downsides and that I would rather that the budget the company allocates to the transistors be used for other things much more relevant, even if it is still FP math (on the GPU, instead of the AVX-512).
Otherwise you want us to just give you more cores (with good single-threaded performance, but no waste like the AVX-512) like AMD did. As such, remember that you recently moved from an Intel processor to an AMD Threadripper processor.
Torvalds said that with this processor his "allmodconfig" tests are three times faster than before. Still compared to the AVX-512, it said that it wants its power limits to be reached with a normal integer code, and not with an AVX-512 power virus that suppresses the maximum frequency (because people ended up using it for memcpy!) .
Finally, Linus Torvalds said that he thinks the AVX-512 is something wrong.
Yes, yes, I am biased. I absolutely disapprove of FP benchmarks and realize that other people deeply care about them. I just think the AVX-512 is precisely the wrong thing to do. It's one of my hobbies. This is a good example of what Intel has done wrong, in part simply by increasing market fragmentation, "he said.
If you want to know more about it, you can consult the following link.